Anthropic Reports 90, 99% Drop in Distillation After New Defenses

Anthropic published a report on October 12, 2026 describing its defenses against distillation attacks, in which third parties harvest large volumes of Claude outputs to train competing models. The company said its countermeasures, including tighter account limits, proxy-network blocks, and a new 'preserved thinking' feature introduced with Claude Fable 5.1 and Claude Opus 5.5, cut data harvested from Claude by 90% to 99%. Preserved thinking prevents API accounts created on or after August 31, 2026 from editing Claude's prior context, making proxy-network extraction harder. Anthropic also blocked specific extraction techniques, such as asking Claude to translate its reasoning into another language or write it out in special tags, and prevented newer models' reasoning from being read by older models with weaker anti-distillation safeguards. The company said it has turned away meaningful revenue and trillions of tokens of paying traffic from suspected distillers. Following Anthropic's measures, Chinese labs reportedly redirected distillation efforts toward OpenAI's GPT-6 Astra, according to The Information. CISA separately documented that multiple China-based AI companies, including StepFun, Z.AI, and Alibaba, conducted systematic distillation of Claude and GPT product lines between late 2025 and mid-2026. Anthropic and OpenAI, along with Google DeepMind, have been sharing threat intelligence through the Frontier Model Forum to coordinate responses. In July 2026, the White House accused Moonshot AI specifically of using an internal distillation platform against Claude Fable 5, a claim Moonshot employees and outside experts publicly disputed as politically motivated and technically implausible given the timeline.